The unprecedented increase of commercial airlines and private jets over the next ten years presents a challenge for air traffic control. Precise flight trajectory prediction is of great significance in air transportation management, which contributes to the decision-making for safe and orderly flights. Existing research and application mainly focus on the sequence generation based on historical trajectories, while the aircraft-aircraft interactions in crowded airspace especially the airspaces near busy airports have been largely ignored. On the other hand, there are distinct characteristics of aerodynamics for different flight phases, and the trajectory may be affected by various uncertainties such as weather and advisories from air traffic controllers. However, there is no literature fully considers all these issues. Therefore, we proposed a phased flight trajectory prediction framework. Multi-source and multi-modal datasets have been analyzed and mined using variants of recurrent neural network (RNN) mixture. To be specific, we first introduce spatio temporal graphs into the low-altitude airway prediction problem, and the motion constraints of an aircraft are embedded to the inference process for reliable forecasting results. In the en-route phase, the dual attention mechanism is employed to adaptively extract much more important features from overall datasets to learn the hidden patterns in dynamical environments. The experimental results demonstrate our proposed framework can outperform state-of-the-art methods for flight trajectory prediction for large passenger/transport airplanes.
The reconstruction of cortical surfaces from brain magnetic resonance imaging (MRI) scans is essential for quantitative analyses of cortical thickness and sulcal morphology. Although traditional and deep learning-based algorithmic pipelines exist for this purpose, they have two major drawbacks: lengthy runtimes of multiple hours (traditional) or intricate post-processing, such as mesh extraction and topology correction (deep learning-based). In this work, we address both of these issues and propose Vox2Cortex, a deep learning-based algorithm that directly yields topologically correct, three-dimensional meshes of the boundaries of the cortex. Vox2Cortex leverages convolutional and graph convolutional neural networks to deform an initial template to the densely folded geometry of the cortex represented by an input MRI scan. We show in extensive experiments on three brain MRI datasets that our meshes are as accurate as the ones reconstructed by state-of-the-art methods in the field, without the need for time- and resource-intensive post-processing. To accurately reconstruct the tightly folded cortex, we work with meshes containing about 168,000 vertices at test time, scaling deep explicit reconstruction methods to a new level.
The emergence of explainability methods has enabled a better comprehension of how deep neural networks operate through concepts that are easily understood and implemented by the end user. While most explainability methods have been designed for traditional deep learning, some have been further developed for geometric deep learning, in which data are predominantly represented as graphs. These representations are regularly derived from medical imaging data, particularly in the field of neuroimaging, in which graphs are used to represent brain structural and functional wiring patterns (brain connectomes) and cortical surface models are used to represent the anatomical structure of the brain. Although explainability techniques have been developed for identifying important vertices (brain areas) and features for graph classification, these methods are still lacking for more complex tasks, such as surface-based modality transfer (or vertex-wise regression). Here, we address the need for surface-based explainability approaches by developing a framework for cortical surface-based deep learning, providing a transparent system for modality transfer tasks. First, we adapted a perturbation-based approach for use with surface data. Then, we applied our perturbation-based method to investigate the key features and vertices used by a geometric deep learning model developed to predict brain function from anatomy directly on a cortical surface model. We show that our explainability framework is not only able to identify important features and their spatial location but that it is also reliable and valid.
This paper introduces a deep learning modular network for side-channel analysis. Our deep learning approach features the capability to exchange part of it (modules) with others networks. We aim to introduce reusable trained modules into side-channel analysis instead of building architectures for each evaluation, reducing the body of work when conducting those. Our experiments demonstrate that our architecture feasibly assesses a side-channel evaluation suggesting that learning transferability is possible with the network we propose in this paper.
Image data labeling is a vital step for deep learning model training. Studies on data labeling have not considered its impact on model performance and only focused on problems such as the curse of big data labeling or labeling tools. Furthermore, it seems clear that errors in labeling have a significant impact and should be fixed. However, in the medical domain, it is hard to ensure proper data labeling. In general, trained engineers are asked to annotate histology images, which causes errors in labeling. The aim of this study is to highlight the impact of data labeling on deep learning models. For that purpose, deep learning models are trained on two different annotations with different levels of expertise. Results show the importance of including expertise in deep learning model development. The impact of data labeling is shown through a case study on the proliferation of biomarker Ki-67 labeling index scoring.
The prior knowledge (a.k.a. priors) integrated into the design of a machine learning system strongly influences its generalization abilities. In the specific context of deep learning, some of these priors are poorly understood as they implicitly emerge from the successful heuristics and tentative approximations of biological brains involved in deep learning design. Through the lens of supervised image classification problems, this thesis investigates the implicit integration of a natural clustering prior composed of three statements: (i) natural images exhibit a rich clustered structure, (ii) image classes are composed of multiple clusters and (iii) each cluster contains examples from a single class. The decomposition of classes into multiple clusters implies that supervised deep learning systems could benefit from unsupervised clustering to define appropriate decision boundaries. Hence, this thesis attempts to identify implicit clustering abilities, mechanisms and hyperparameters in deep learning systems and evaluate their relevance for explaining the generalization abilities of these systems. We do so through an extensive empirical study of the training dynamics as well as the neuron- and layer-level representations of deep neural networks. The resulting collection of experiments provides preliminary evidence for the relevance of the natural clustering prior for understanding deep learning.
This paper presents primarily two Euclidean embeddings of the quotient space generated by matrices that are identified modulo arbitrary row permutations. The original application is in deep learning on graphs where the learning task is invariant to node relabeling. Two embedding schemes are introduced, one based on sorting and the other based on algebras of multivariate polynomials. While both embeddings exhibit a computational complexity exponential in problem size, the sorting based embedding is globally bi-Lipschitz and admits a low dimensional target space. Additionally, an almost everywhere injective scheme can be implemented with minimal redundancy and low computational cost. In turn, this proves that almost any classifier can be implemented with an arbitrary small loss of performance. Numerical experiments are carried out on two data sets, a chemical compound data set (QM9) and a proteins data set (PROTEINS).
Training very deep neural networks is still an extremely challenging task. The common solution is to use shortcut connections and normalization layers, which are both crucial ingredients in the popular ResNet architecture. However, there is strong evidence to suggest that ResNets behave more like ensembles of shallower networks than truly deep ones. Recently, it was shown that deep vanilla networks (i.e. networks without normalization layers or shortcut connections) can be trained as fast as ResNets by applying certain transformations to their activation functions. However, this method (called Deep Kernel Shaping) isn't fully compatible with ReLUs, and produces networks that overfit significantly more than ResNets on ImageNet. In this work, we rectify this situation by developing a new type of transformation that is fully compatible with a variant of ReLUs -- Leaky ReLUs. We show in experiments that our method, which introduces negligible extra computational cost, achieves validation accuracies with deep vanilla networks that are competitive with ResNets (of the same width/depth), and significantly higher than those obtained with the Edge of Chaos (EOC) method. And unlike with EOC, the validation accuracies we obtain do not get worse with depth.
Autonomous driving requires a detailed understanding of complex driving scenes. The redundancy and complementarity of the vehicle's sensors provide an accurate and robust comprehension of the environment, thereby increasing the level of performance and safety. This thesis focuses the on automotive RADAR, which is a low-cost active sensor measuring properties of surrounding objects, including their relative speed, and has the key advantage of not being impacted by adverse weather conditions. With the rapid progress of deep learning and the availability of public driving datasets, the perception ability of vision-based driving systems has considerably improved. The RADAR sensor is seldom used for scene understanding due to its poor angular resolution, the size, noise, and complexity of RADAR raw data as well as the lack of available datasets. This thesis proposes an extensive study of RADAR scene understanding, from the construction of an annotated dataset to the conception of adapted deep learning architectures. First, this thesis details approaches to tackle the current lack of data. A simple simulation as well as generative methods for creating annotated data will be presented. It will also describe the CARRADA dataset, composed of synchronised camera and RADAR data with a semi-automatic annotation method. This thesis then present a proposed set of deep learning architectures with their associated loss functions for RADAR semantic segmentation. It also introduces a method to open up research into the fusion of LiDAR and RADAR sensors for scene understanding. Finally, this thesis exposes a collaborative contribution, the RADIal dataset with synchronised High-Definition (HD) RADAR, LiDAR and camera. A deep learning architecture is also proposed to estimate the RADAR signal processing pipeline while performing multitask learning for object detection and free driving space segmentation.
We propose an unsupervised deep learning based method to estimate depth from focal stack camera images. On the NYU-v2 dataset, our method achieves much better depth estimation accuracy compared to single-image based methods.
With the number of vehicles continuously increasing, parking monitoring and analysis are becoming a substantial feature of modern cities. In this study, we present a methodology to monitor car parking areas and to analyze their occupancy in real-time. The solution is based on a combination between image analysis and deep learning techniques. It incorporates four building blocks put inside a pipeline: vehicle detection, vehicle tracking, manual annotation of parking slots, and occupancy estimation using the Ray Tracing algorithm. The aim of this methodology is to optimize the use of parking areas and to reduce the time wasted by daily drivers to find the right parking slot for their cars. Also, it helps to better manage the space of the parking areas and to discover misuse cases. A demonstration of the provided solution is shown in the following video link: https://www.youtube.com/watch?v=KbAt8zT14Tc.
Many recent deep learning-based solutions have widely adopted the attention-based mechanism in various tasks of the NLP discipline. However, the inherent characteristics of deep learning models and the flexibility of the attention mechanism increase the models' complexity, thus leading to challenges in model explainability. In this paper, to address this challenge, we propose a novel practical framework by utilizing a two-tier attention architecture to decouple the complexity of explanation and the decision-making process. We apply it in the context of a news article classification task. The experiments on two large-scaled news corpora demonstrate that the proposed model can achieve competitive performance with many state-of-the-art alternatives and illustrate its appropriateness from an explainability perspective.
The energy market relies on forecasting capabilities of both demand and power generation that need to be kept in dynamic balance. Today, when it comes to renewable energy generation, such decisions are increasingly made in a liberalized electricity market environment, where future power generation must be offered through contracts and auction mechanisms, hence based on forecasts. The increased share of highly intermittent power generation from renewable energy sources increases the uncertainty about the expected future power generation. Point forecast does not account for such uncertainties. To account for these uncertainties, it is possible to make probabilistic forecasts. This work first presents important concepts and approaches concerning probabilistic forecasts with deep learning. Then, deep learning models are used to make probabilistic forecasts of day-ahead power generation from a wind power plant located in Northern Norway. The performance in terms of obtained quality of the prediction intervals is compared for different deep learning models and sets of covariates. The findings show that the accuracy of the predictions improves when historical data on measured weather and numerical weather predictions (NWPs) were included as exogenous variables. This allows the model to auto-correct systematic biases in the NWPs using the historical measurement data. Using only NWPs, or only measured weather as exogenous variables, worse prediction performances were obtained.
This paper combines the power of deep-learning with the generalizability of physics-based features, to present an advanced method for seismic discrimination between earthquakes and explosions. The proposed method contains two branches: a deep learning branch operating directly on seismic waveforms or spectrograms, and a second branch operating on physics-based parametric features. These features are high-frequency P/S amplitude ratios and the difference between local magnitude (ML) and coda duration magnitude (MC). The combination achieves better generalization performance when applied to new regions than models that are developed solely with deep learning. We also examined which parts of the waveform data dominate deep learning decisions (i.e., via Grad-CAM). Such visualization provides a window into the black-box nature of the machine-learning models and offers new insight into how the deep learning derived models use data to make the decisions.
Food is essential for human survival, and people always try to taste different types of delicious recipes. Frequently, people choose food ingredients without even knowing their names or pick up some food ingredients that are not obvious to them from a grocery store. Knowing which ingredients can be mixed to make a delicious food recipe is essential. Selecting the right recipe by choosing a list of ingredients is very difficult for a beginner cook. However, it can be a problem even for experts. One such example is recognising objects through image processing. Although this process is complex due to different food ingredients, traditional approaches will lead to an inaccuracy rate. These problems can be solved by machine learning and deep learning approaches. In this paper, we implemented a model for food ingredients recognition and designed an algorithm for recommending recipes based on recognised ingredients. We made a custom dataset consisting of 9856 images belonging to 32 different food ingredients classes. Convolution Neural Network (CNN) model was used to identify food ingredients, and for recipe recommendations, we have used machine learning. We achieved an accuracy of 94 percent, which is quite impressive.
The rapid advancement of deep learning has significantly boomed the development of photorealistic style transfer. In this review, we reviewed the development of photorealistic style transfer starting from artistic style transfer and the contribution of traditional image processing techniques on photorealistic style transfer, including some work that had been completed in the Multimedia lab at the University of Alberta. Many techniques were discussed in this review. However, our focus is on VGG-based techniques, whitening and coloring transform (WCTs) based techniques, the combination of deep learning with traditional image processing techniques.
Point cloud denoising aims to restore clean point clouds from raw observations corrupted by noise and outliers while preserving the fine-grained details. We present a novel deep learning-based denoising model, that incorporates normalizing flows and noise disentanglement techniques to achieve high denoising accuracy. Unlike existing works that extract features of point clouds for point-wise correction, we formulate the denoising process from the perspective of distribution learning and feature disentanglement. By considering noisy point clouds as a joint distribution of clean points and noise, the denoised results can be derived from disentangling the noise counterpart from latent point representation, and the mapping between Euclidean and latent spaces is modeled by normalizing flows. We evaluate our method on synthesized 3D models and real-world datasets with various noise settings. Qualitative and quantitative results show that our method outperforms previous state-of-the-art deep learning-based approaches.
Solar activity has significant impacts on human activities and health. One most commonly used measure of solar activity is the sunspot number. This paper compares three important non-deep learning models, four popular deep learning models, and their five ensemble models in forecasting sunspot numbers. In particular, we propose an ensemble model called XGBoost-DL, which uses XGBoost as a two-level nonlinear ensemble method to combine the deep learning models. Our XGBoost-DL achieves the best forecasting performance (RMSE = 25.70 and MAE = 19.82) in the comparison, outperforming the best non-deep learning model SARIMA (RMSE = 54.11 and MAE = 45.51), the best deep learning model Informer (RMSE = 29.90 and MAE = 22.35) and the NASA's forecast (RMSE = 48.38 and MAE = 38.45). Our XGBoost-DL forecasts a peak sunspot number of 133.47 in May 2025 for Solar Cycle 25 and 164.62 in November 2035 for Solar Cycle 26, similar to but later than the NASA's at 137.7 in October 2024 and 161.2 in December 2034. An open-source Python package of our XGBoost-DL for the sunspot number prediction is available at https://github.com/yd1008/ts_ensemble_sunspot.
Retinal implants have the potential to treat incurable blindness, yet the quality of the artificial vision they produce is still rudimentary. An outstanding challenge is identifying electrode activation patterns that lead to intelligible visual percepts (phosphenes). Here we propose a PSE based on CNN that is trained in an end-to-end fashion to predict the electrode activation patterns required to produce a desired visual percept. We demonstrate the effectiveness of the encoder on MNIST using a psychophysically validated phosphene model tailored to individual retinal implant users. The present work constitutes an essential first step towards improving the quality of the artificial vision provided by retinal implants.
The Benes filter is a well-known continuous-time stochastic filtering model in one dimension that has the advantage of being explicitly solvable. From an evolution equation point of view, the Benes filter is also the solution of the filtering equations given a particular set of coefficient functions. In general, the filtering stochastic partial differential equations (SPDE) arise as the evolution equations for the conditional distribution of an underlying signal given partial, and possibly noisy, observations. Their numerical approximation presents a central issue for theoreticians and practitioners alike, who are actively seeking accurate and fast methods, especially for such high-dimensional settings as numerical weather prediction, for example. In this paper we present a brief study of a new numerical method based on the mesh-free neural network representation of the density of the solution of the Benes model achieved by deep learning. Based on the classical SPDE splitting method, our algorithm includes a recursive normalisation procedure to recover the normalised conditional distribution of the signal process. Within the analytically tractable setting of the Benes filter, we discuss the role of nonlinearity in the filtering model equations for the choice of the domain of the neural network. Further we present the first study of the neural network method with an adaptive domain for the Benes model.